def test_perform_feature_scaling_false_regression():
np.random.seed(42)
df_boston_train, df_boston_test = utils.get_boston_regression_dataset()
column_descriptions = {
'MEDV': 'output'
, 'CHAS': 'categorical'
}
ml_predictor = Predictor(type_of_estimator='regressor', column_descriptions=column_descriptions)
ml_predictor.train(df_boston_train, perform_feature_scaling=False, model_names=['DeepLearningRegressor'])
file_name = ml_predictor.save(str(random.random()))
saved_ml_pipeline = utils_models.load_keras_model(file_name)
# with open(file_name, 'rb') as read_file:
# saved_ml_pipeline = dill.load(read_file)
os.remove(file_name)
test_score = saved_ml_pipeline.score(df_boston_test, df_boston_test.MEDV)
print('test_score')
print(test_score)
assert -24 < test_score < -2.8
def test_perform_feature_scaling_false_regression(model_name=None):
np.random.seed(0)
df_boston_train, df_boston_test = utils.get_boston_regression_dataset()
column_descriptions = {
'MEDV': 'output'
, 'CHAS': 'categorical'
}
ml_predictor = Predictor(type_of_estimator='regressor', column_descriptions=column_descriptions)
ml_predictor.train(df_boston_train, perform_feature_scaling=False, model_names=model_name)
test_score = ml_predictor.score(df_boston_test, df_boston_test.MEDV)
print('test_score')
print(test_score)
lower_bound = -3.2
if model_name == 'DeepLearningRegressor':
lower_bound = -8.8
if model_name == 'LGBMRegressor':
lower_bound = -4.95
assert lower_bound < test_score < -2.8
def test_is_backwards_compatible_with_models_trained_using_1_9_6():
np.random.seed(0)
df_boston_train, df_boston_test = utils.get_boston_regression_dataset()
with open(os.path.join('tests', 'trained_ml_model_v_1_9_6.dill'),
'rb') as read_file:
saved_ml_pipeline = dill.load(read_file)
df_boston_test_dictionaries = df_boston_test.to_dict('records')
# 1. make sure the accuracy is the same
predictions = []
for row in df_boston_test_dictionaries:
predictions.append(saved_ml_pipeline.predict(row))
print('predictions')
print(predictions)
print('predictions[0]')
print(predictions[0])
print('type(predictions)')
print(type(predictions))
first_score = utils.calculate_rmse(df_boston_test.MEDV, predictions)
print('first_score')
print(first_score)
# Make sure our score is good, but not unreasonably good
assert -2.8 < first_score < -2.1
# 2. make sure the speed is reasonable (do it a few extra times)
data_length = len(df_boston_test_dictionaries)
start_time = datetime.datetime.now()
for idx in range(1000):
row_num = idx % data_length
saved_ml_pipeline.predict(df_boston_test_dictionaries[row_num])
end_time = datetime.datetime.now()
duration = end_time - start_time
print('duration.total_seconds()')
print(duration.total_seconds())
# It's very difficult to set a benchmark for speed that will work across all machines.
# On my 2013 bottom of the line 15" MacBook Pro, this runs in about 0.8 seconds for 1000 predictions
# That's about 1 millisecond per prediction
# Assuming we might be running on a test box that's pretty weak, multiply by 3
# Also make sure we're not running unreasonably quickly
assert 0.1 < duration.total_seconds() / 1.0 < 15
# 3. make sure we're not modifying the dictionaries (the score is the same after running a few experiments as it is the first time)
predictions = []
for row in df_boston_test_dictionaries:
predictions.append(saved_ml_pipeline.predict(row))
second_score = utils.calculate_rmse(df_boston_test.MEDV, predictions)
print('second_score')
print(second_score)
assert -2.8 < second_score < -2.1
def test_input_df_unmodified():
np.random.seed(42)
df_boston_train, df_boston_test = utils.get_boston_regression_dataset()
column_descriptions = {
'MEDV': 'output'
, 'CHAS': 'categorical'
}
ml_predictor = Predictor(type_of_estimator='regressor', column_descriptions=column_descriptions)
df_shape = df_boston_train.shape
ml_predictor.train(df_boston_train)
training_shape = df_boston_train.shape
assert training_shape[0] == df_shape[0]
assert training_shape[1] == df_shape[1]
test_score = ml_predictor.score(df_boston_test, df_boston_test.MEDV)
print('test_score')
print(test_score)
assert -3.35 < test_score < -2.8
def categorical_ensembling_regression(model_name=None):
np.random.seed(0)
df_boston_train, df_boston_test = utils.get_boston_regression_dataset()
column_descriptions = {'MEDV': 'output', 'CHAS': 'categorical'}
ml_predictor = Predictor(type_of_estimator='regressor',
column_descriptions=column_descriptions)
ml_predictor.train_categorical_ensemble(df_boston_train,
perform_feature_selection=True,
model_names=model_name,
categorical_column='CHAS')
test_score = ml_predictor.score(df_boston_test, df_boston_test.MEDV)
print('test_score')
print(test_score)
# Bumping this up since without these features our score drops
lower_bound = -4.0
if model_name == 'DeepLearningRegressor':
lower_bound = -19
if model_name == 'LGBMRegressor':
lower_bound = -4.95
assert lower_bound < test_score < -2.8
def ensemble_regressor_basic_test():
np.random.seed(0)
df_boston_train, df_boston_test = utils.get_boston_regression_dataset()
column_descriptions = {
'MEDV': 'output'
, 'CHAS': 'categorical'
}
ensemble_config = [
{
'model_name': 'LGBMRegressor'
}
, {
'model_name': 'RandomForestRegressor'
}
]
ml_predictor = Predictor(type_of_estimator='regressor', column_descriptions=column_descriptions)
ml_predictor.train(df_boston_train, ensemble_config=None)
test_score = ml_predictor.score(df_boston_test, df_boston_test.MEDV)
print('test_score')
print(test_score)
assert -3.0 < test_score < -2.8
def test_prediction_intervals_actually_work():
np.random.seed(0)
df_boston_train, df_boston_test = utils.get_boston_regression_dataset()
column_descriptions = {'MEDV': 'output', 'CHAS': 'categorical'}
ml_predictor = Predictor(type_of_estimator='regressor',
column_descriptions=column_descriptions)
ml_predictor.train(df_boston_train, predict_intervals=[0.05, 0.95])
df_boston_test = df_boston_test.reset_index(drop=True)
intervals = ml_predictor.predict_intervals(df_boston_test)
actuals = df_boston_test.MEDV
count_under = 0
count_over = 0
# print(intervals)
for idx, row in intervals.iterrows():
actual = actuals.iloc[idx]
if actual < row['interval_0.05']:
count_under += 1
if actual > row['interval_0.95']:
count_over += 1
len_intervals = len(intervals)
pct_under = count_under * 1.0 / len_intervals
pct_over = count_over * 1.0 / len_intervals
# There's a decent bit of noise since this is such a small dataset
assert pct_under < 0.15
assert pct_over < 0.1
def test_select_from_multiple_regression_models_using_X_test_and_y_test():
np.random.seed(0)
df_boston_train, df_boston_test = utils.get_boston_regression_dataset()
column_descriptions = {'MEDV': 'output', 'CHAS': 'categorical'}
ml_predictor = Predictor(type_of_estimator='regressor',
column_descriptions=column_descriptions)
ml_predictor.train(df_boston_train,
model_names=[
'LinearRegression', 'RandomForestRegressor',
'Ridge', 'GradientBoostingRegressor',
'ExtraTreesRegressor', 'AdaBoostRegressor',
'SGDRegressor', 'PassiveAggressiveRegressor'
],
X_test=df_boston_test,
y_test=df_boston_test.MEDV)
test_score = ml_predictor.score(df_boston_test, df_boston_test.MEDV)
print('test_score')
print(test_score)
# Due to the small sample size of this test set, GSCV may sometimes pick ExtraTreesRegressor as the best model, just very slightly beating out GradientBoostingRegressor.
# ExtraTrees doesn't generalize as well, however, scoring a mere -3.20x something or other, and narrowly missing our cutoff from above.
# Given that is is only an issue when running on tiny toy datasets, I'm not concerned for the use cases I intend to support, and thus, am bumping up the upper bound on our error metric ever so slightly
assert -3.25 < test_score < -2.8
def test_score_uncertainty():
np.random.seed(0)
df_boston_train, df_boston_test = utils.get_boston_regression_dataset()
column_descriptions = {'MEDV': 'output', 'CHAS': 'categorical'}
df_boston_train, uncertainty_data = train_test_split(df_boston_train,
test_size=0.5)
ml_predictor = Predictor(type_of_estimator='regressor',
column_descriptions=column_descriptions)
ml_predictor.train(df_boston_train,
perform_feature_selection=True,
train_uncertainty_model=True,
uncertainty_data=uncertainty_data)
uncertainty_score = ml_predictor.score_uncertainty(df_boston_test,
df_boston_test.MEDV)
print('uncertainty_score')
print(uncertainty_score)
assert uncertainty_score > -0.2
def optimize_final_model_regression(model_name=None):
np.random.seed(0)
df_boston_train, df_boston_test = utils.get_boston_regression_dataset()
column_descriptions = {'MEDV': 'output', 'CHAS': 'categorical'}
ml_predictor = Predictor(type_of_estimator='regressor',
column_descriptions=column_descriptions)
ml_predictor.train(df_boston_train,
optimize_final_model=True,
model_names=model_name)
test_score = ml_predictor.score(df_boston_test, df_boston_test.MEDV)
print('test_score')
print(test_score)
# the random seed gets a score of -3.21 on python 3.5
# There's a ton of noise here, due to small sample sizes
lower_bound = -3.4
if model_name == 'DeepLearningRegressor':
lower_bound = -20
if model_name == 'LGBMRegressor':
lower_bound = -5.5
if model_name == 'GradientBoostingRegressor':
lower_bound = -3.5
assert lower_bound < test_score < -2.8
def test_prediction_intervals_actually_work():
np.random.seed(0)
df_boston_train, df_boston_test = utils.get_boston_regression_dataset()
column_descriptions = {'MEDV': 'output', 'CHAS': 'categorical'}
df_boston_train, uncertainty_data = train_test_split(df_boston_train,
test_size=0.5)
ml_predictor = Predictor(type_of_estimator='regressor',
column_descriptions=column_descriptions)
ml_predictor.train(df_boston_train, predict_intervals=True)
intervals = ml_predictor.predict_intervals(df_boston_test)
count_under = 0
count_over = 0
for row in intervals:
if row[0] < row[1]:
count_under += 1
if row[0] > row[3]:
count_over += 1
len_intervals = len(intervals)
pct_under = count_under * 1.0 / len_intervals
pct_over = count_over * 1.0 / len_intervals
# There's a decent bit of noise since this is such a small dataset
assert pct_under < 0.1
assert pct_over < 0.1
def test_all_algos_regression():
# a random seed of 42 has ExtraTreesRegressor getting the best CV score, and that model doesn't generalize as well as GradientBoostingRegressor.
np.random.seed(0)
df_boston_train, df_boston_test = utils.get_boston_regression_dataset()
column_descriptions = {'MEDV': 'output', 'CHAS': 'categorical'}
ml_predictor = Predictor(type_of_estimator='regressor',
column_descriptions=column_descriptions)
ml_predictor.train(
df_boston_train,
model_names=[
'LinearRegression', 'RandomForestRegressor', 'Ridge',
'GradientBoostingRegressor', 'AdaBoostRegressor', 'SGDRegressor',
'PassiveAggressiveRegressor', 'Lasso', 'LassoLars', 'ElasticNet',
'OrthogonalMatchingPursuit', 'BayesianRidge', 'ARDRegression',
'MiniBatchKMeans', 'DeepLearningRegressor', 'LGBMRegressor',
'XGBClassifier', 'LinearSVR', 'CatBoostRegressor'
])
test_score = ml_predictor.score(df_boston_test, df_boston_test.MEDV)
print('test_score')
print(test_score)
assert -3.4 < test_score < -2.8
def test_calibrate_uncertainty():
np.random.seed(0)
df_boston_train, df_boston_test = utils.get_boston_regression_dataset()
column_descriptions = {'MEDV': 'output', 'CHAS': 'categorical'}
df_boston_train, uncertainty_data = train_test_split(df_boston_train,
test_size=0.5)
uncertainty_data, uncertainty_calibration_data = train_test_split(
uncertainty_data, test_size=0.5)
ml_predictor = Predictor(type_of_estimator='regressor',
column_descriptions=column_descriptions)
uncertainty_calibration_settings = {
'num_buckets': 3,
'percentiles': [25, 50, 75]
}
ml_predictor.train(
df_boston_train,
perform_feature_selection=True,
train_uncertainty_model=True,
uncertainty_data=uncertainty_data,
calibrate_uncertainty=True,
uncertainty_calibration_settings=uncertainty_calibration_settings,
uncertainty_calibration_data=uncertainty_calibration_data)
uncertainty_score = ml_predictor.predict_uncertainty(df_boston_test)
assert 'percentile_25_delta' in list(uncertainty_score.columns)
assert 'percentile_50_delta' in list(uncertainty_score.columns)
assert 'percentile_75_delta' in list(uncertainty_score.columns)
assert 'bucket_num' in list(uncertainty_score.columns)
def test_all_algos_regression():
# a random seed of 42 has ExtraTreesRegressor getting the best CV score, and that model doesn't generalize as well as GradientBoostingRegressor.
np.random.seed(0)
df_boston_train, df_boston_test = utils.get_boston_regression_dataset()
column_descriptions = {'MEDV': 'output', 'CHAS': 'categorical'}
ml_predictor = Predictor(type_of_estimator='regressor',
column_descriptions=column_descriptions)
ml_predictor.train(df_boston_train,
model_names=[
'LinearRegression', 'RandomForestRegressor',
'Ridge', 'GradientBoostingRegressor',
'ExtraTreesRegressor', 'AdaBoostRegressor',
'SGDRegressor', 'PassiveAggressiveRegressor'
])
test_score = ml_predictor.score(df_boston_test, df_boston_test.MEDV)
print('test_score')
print(test_score)
assert -3.25 < test_score < -2.8
def test_regression():
df_boston_train, df_boston_test = utils.get_boston_regression_dataset()
ml_predictor = utils.train_basic_regressor(df_boston_train)
test_score = ml_predictor.score(df_boston_test,
df_boston_test.MEDV,
verbose=0)
# Currently, we expect to get a score of -3.09
# Make sure our score is good, but not unreasonably good
assert -3.2 < test_score < -2.8
def test_getting_single_predictions_regression():
np.random.seed(0)
df_boston_train, df_boston_test = utils.get_boston_regression_dataset()
ml_predictor = utils.train_basic_regressor(df_boston_train)
file_name = ml_predictor.save(str(random.random()))
with open(file_name, 'rb') as read_file:
saved_ml_pipeline = dill.load(read_file)
os.remove(file_name)
df_boston_test_dictionaries = df_boston_test.to_dict('records')
# 1. make sure the accuracy is the same
predictions = []
for row in df_boston_test_dictionaries:
predictions.append(saved_ml_pipeline.predict(row))
first_score = utils.calculate_rmse(df_boston_test.MEDV, predictions)
print('first_score')
print(first_score)
# Make sure our score is good, but not unreasonably good
assert -3.2 < first_score < -2.8
# 2. make sure the speed is reasonable (do it a few extra times)
data_length = len(df_boston_test_dictionaries)
start_time = datetime.datetime.now()
for idx in range(1000):
row_num = idx % data_length
saved_ml_pipeline.predict(df_boston_test_dictionaries[row_num])
end_time = datetime.datetime.now()
duration = end_time - start_time
print('duration.total_seconds()')
print(duration.total_seconds())
# It's very difficult to set a benchmark for speed that will work across all machines.
# On my 2013 bottom of the line 15" MacBook Pro, this runs in about 0.8 seconds for 1000 predictions
# That's about 1 millisecond per prediction
# Assuming we might be running on a test box that's pretty weak, multiply by 3
# Also make sure we're not running unreasonably quickly
assert 0.2 < duration.total_seconds() / 1.0 < 3
# 3. make sure we're not modifying the dictionaries (the score is the same after running a few experiments as it is the first time)
predictions = []
for row in df_boston_test_dictionaries:
predictions.append(saved_ml_pipeline.predict(row))
second_score = utils.calculate_rmse(df_boston_test.MEDV, predictions)
print('second_score')
print(second_score)
# Make sure our score is good, but not unreasonably good
assert -3.2 < second_score < -2.8
def test_saving_trained_pipeline_regression():
df_boston_train, df_boston_test = utils.get_boston_regression_dataset()
ml_predictor = utils.train_basic_regressor(df_boston_train)
file_name = ml_predictor.save()
with open(file_name, 'rb') as read_file:
saved_ml_pipeline = dill.load(read_file)
test_score = saved_ml_pipeline.score(df_boston_test, df_boston_test.MEDV)
# Make sure our score is good, but not unreasonably good
assert -3.2 < test_score < -2.8
def test_predict_uncertainty_true():
np.random.seed(0)
df_boston_train, df_boston_test = utils.get_boston_regression_dataset()
column_descriptions = {'MEDV': 'output', 'CHAS': 'categorical'}
df_boston_train, uncertainty_data = train_test_split(df_boston_train,
test_size=0.5)
ml_predictor = Predictor(type_of_estimator='regressor',
column_descriptions=column_descriptions)
ml_predictor.train(df_boston_train, predict_intervals=True)
intervals = ml_predictor.predict_intervals(df_boston_test)
assert isinstance(intervals, pd.DataFrame)
assert intervals.shape[0] == df_boston_test.shape[0]
result_list = ml_predictor.predict_intervals(df_boston_test,
return_type='list')
assert isinstance(result_list, list)
assert len(result_list) == df_boston_test.shape[0]
for idx, row in enumerate(result_list):
assert isinstance(row, list)
assert len(row) == 3
singles = df_boston_test.head().to_dict('records')
for row in singles:
result = ml_predictor.predict_intervals(row)
assert isinstance(result, dict)
assert 'prediction' in result
assert 'interval_0.05' in result
assert 'interval_0.95' in result
for row in singles:
result = ml_predictor.predict_intervals(row, return_type='list')
assert isinstance(result, list)
assert len(result) == 3
df_intervals = ml_predictor.predict_intervals(df_boston_test,
return_type='df')
assert isinstance(df_intervals, pd.DataFrame)
try:
ml_predictor.predict_intervals(df_boston_test,
return_type='this will not work')
assert False
except ValueError:
assert True
def test_predict_intervals_should_fail_if_not_trained():
np.random.seed(0)
df_boston_train, df_boston_test = utils.get_boston_regression_dataset()
column_descriptions = {'MEDV': 'output', 'CHAS': 'categorical'}
ml_predictor = Predictor(type_of_estimator='regressor',
column_descriptions=column_descriptions)
ml_predictor.train(df_boston_train)
try:
intervals = ml_predictor.predict_intervals(df_boston_test)
assert False
except ValueError:
assert True
def test_prediction_intervals_lets_the_user_specify_number_of_intervals():
np.random.seed(0)
df_boston_train, df_boston_test = utils.get_boston_regression_dataset()
column_descriptions = {'MEDV': 'output', 'CHAS': 'categorical'}
ml_predictor = Predictor(type_of_estimator='regressor',
column_descriptions=column_descriptions)
ml_predictor.train(df_boston_train,
predict_intervals=True,
prediction_intervals=[.2])
intervals = ml_predictor.predict_intervals(df_boston_test,
return_type='list')
assert len(intervals[0]) == 2
def test_all_algos_regression():
df_boston_train, df_boston_test = utils.get_boston_regression_dataset()
column_descriptions = {
'MEDV': 'output'
, 'CHAS': 'categorical'
}
ml_predictor = Predictor(type_of_estimator='regressor', column_descriptions=column_descriptions)
ml_predictor.train(df_boston_train, model_names=['LinearRegression', 'RandomForestRegressor', 'Ridge', 'GradientBoostingRegressor', 'ExtraTreesRegressor', 'AdaBoostRegressor', 'SGDRegressor', 'PassiveAggressiveRegressor'])
test_score = ml_predictor.score(df_boston_test, df_boston_test.MEDV)
print('test_score')
print(test_score)
assert -3.25 < test_score < -2.8
def test_compute_power_1_regression():
df_boston_train, df_boston_test = utils.get_boston_regression_dataset()
column_descriptions = {
'MEDV': 'output'
, 'CHAS': 'categorical'
}
ml_predictor = Predictor(type_of_estimator='regressor', column_descriptions=column_descriptions)
ml_predictor.train(df_boston_train, compute_power=1)
test_score = ml_predictor.score(df_boston_test, df_boston_test.MEDV)
print('test_score')
print(test_score)
assert -3.2 < test_score < -2.8
def test_compare_all_models_regression():
np.random.seed(0)
df_boston_train, df_boston_test = utils.get_boston_regression_dataset()
column_descriptions = {'MEDV': 'output', 'CHAS': 'categorical'}
ml_predictor = Predictor(type_of_estimator='regressor',
column_descriptions=column_descriptions)
ml_predictor.train(df_boston_train, compare_all_models=True)
test_score = ml_predictor.score(df_boston_test, df_boston_test.MEDV)
print('test_score')
print(test_score)
assert -3.2 < test_score < -2.8
def test_perform_feature_selection_false_regression():
np.random.seed(0)
df_boston_train, df_boston_test = utils.get_boston_regression_dataset()
column_descriptions = {'MEDV': 'output', 'CHAS': 'categorical'}
ml_predictor = Predictor(type_of_estimator='regressor',
column_descriptions=column_descriptions)
ml_predictor.train(df_boston_train, perform_feature_selection=False)
test_score = ml_predictor.score(df_boston_test, df_boston_test.MEDV)
print('test_score')
print(test_score)
assert -3.2 < test_score < -2.8
def test_predict_uncertainty_returns_pandas_DataFrame_for_more_than_one_value():
np.random.seed(0)
df_boston_train, df_boston_test = utils.get_boston_regression_dataset()
column_descriptions = {
'MEDV': 'output'
, 'CHAS': 'categorical'
}
df_boston_train, uncertainty_data = train_test_split(df_boston_train, test_size=0.5)
ml_predictor = Predictor(type_of_estimator='regressor', column_descriptions=column_descriptions)
ml_predictor.train(df_boston_train, perform_feature_selection=True, train_uncertainty_model=True, uncertainty_data=uncertainty_data)
uncertainties = ml_predictor.predict_uncertainty(df_boston_test)
assert isinstance(uncertainties, pd.DataFrame)
def test_perform_feature_selection_true_regression():
np.random.seed(0)
df_boston_train, df_boston_test = utils.get_boston_regression_dataset()
column_descriptions = {'MEDV': 'output', 'CHAS': 'categorical'}
ml_predictor = Predictor(type_of_estimator='regressor',
column_descriptions=column_descriptions)
ml_predictor.train(df_boston_train, perform_feature_selection=True)
test_score = ml_predictor.score(df_boston_test, df_boston_test.MEDV)
print('test_score')
print(test_score)
# Bumping this up since without these features our score drops
assert -4.0 < test_score < -2.8
def test_optimize_final_model_regression():
np.random.seed(0)
df_boston_train, df_boston_test = utils.get_boston_regression_dataset()
column_descriptions = {'MEDV': 'output', 'CHAS': 'categorical'}
ml_predictor = Predictor(type_of_estimator='regressor',
column_descriptions=column_descriptions)
ml_predictor.train(df_boston_train, optimize_final_model=True)
test_score = ml_predictor.score(df_boston_test, df_boston_test.MEDV)
print('test_score')
print(test_score)
# the random seed gets a score of -3.21 on python 3.5
assert -3.25 < test_score < -2.8
def test_compute_power_1_regression():
np.random.seed(42)
df_boston_train, df_boston_test = utils.get_boston_regression_dataset()
column_descriptions = {'MEDV': 'output', 'CHAS': 'categorical'}
ml_predictor = Predictor(type_of_estimator='regressor',
column_descriptions=column_descriptions)
ml_predictor.train(df_boston_train,
compute_power=1,
model_names=['LGBMRegressor'])
test_score = ml_predictor.score(df_boston_test, df_boston_test.MEDV)
print('test_score')
print(test_score)
assert -9.5 < test_score < -2.8
def test_perform_feature_scaling_true_regression():
np.random.seed(42)
df_boston_train, df_boston_test = utils.get_boston_regression_dataset()
column_descriptions = {
'MEDV': 'output'
, 'CHAS': 'categorical'
}
ml_predictor = Predictor(type_of_estimator='regressor', column_descriptions=column_descriptions)
ml_predictor.train(df_boston_train, perform_feature_scaling=True, model_names=['DeepLearningRegressor'])
test_score = ml_predictor.score(df_boston_test, df_boston_test.MEDV)
print('test_score')
print(test_score)
assert -24 < test_score < -2.8
def test_predict_uncertainty_returns_dict_for_one_value():
np.random.seed(0)
df_boston_train, df_boston_test = utils.get_boston_regression_dataset()
column_descriptions = {
'MEDV': 'output'
, 'CHAS': 'categorical'
}
df_boston_train, uncertainty_data = train_test_split(df_boston_train, test_size=0.5)
ml_predictor = Predictor(type_of_estimator='regressor', column_descriptions=column_descriptions)
ml_predictor.train(df_boston_train, perform_feature_selection=True, train_uncertainty_model=True, uncertainty_data=uncertainty_data)
test_list = df_boston_test.to_dict('records')
for item in test_list:
prediction = ml_predictor.predict_uncertainty(item)
assert isinstance(prediction, dict)